The paper presents a novel semi-supervised change detection approach called C2F-SemiCD, which consists of two key components:
C2FNet: A coarse-to-fine change detection network with a multi-scale attention mechanism. C2FNet gradually extracts change features from coarse to fine granularity through various modules, including multi-scale feature fusion, channel attention, spatial attention, global context, feature refinement, and feature aggregation.
Semi-supervised learning: C2F-SemiCD employs the mean teacher method for semi-supervised learning, where the teacher model generates pseudo-labels to guide the training of the student model (based on C2FNet). This allows the model to effectively leverage both labeled and unlabeled data.
The authors conduct extensive experiments on three prominent change detection datasets (GoogleGZ-CD, WHU-CD, LEVIR-CD) and perform detailed ablation studies. The results demonstrate that C2F-SemiCD significantly outperforms state-of-the-art supervised and semi-supervised change detection methods, especially when the proportion of labeled data is small (e.g., 5-30%). The proposed approach can effectively extract change features and achieve high-precision change detection, even with limited labeled samples.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Chengxi Han,... klo arxiv.org 04-23-2024
https://arxiv.org/pdf/2404.13838.pdfSyvällisempiä Kysymyksiä